mirror of https://github.com/kubeflow/examples.git
new PR for XGBoost due to problems with history rewrite
This commit is contained in:
Puneith Kaul
parent
e6b6730650
commit
ecc1aab0e4
|
|
@ -0,0 +1,37 @@
|
|||
# Copyright 2018 Google Inc.
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# https://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# Unless required by applicable law or agreed to in writing, software
|
||||
# distributed under the License is distributed on an "AS IS" BASIS,
|
||||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
|
||||
# Use an official Ubuntu 18.04 as parent image
|
||||
FROM ubuntu:18.04
|
||||
|
||||
# Install python 2.7
|
||||
FROM python:2.7
|
||||
|
||||
RUN apt-get update
|
||||
RUN apt-get install -y git make g++
|
||||
|
||||
# Build XGBoost
|
||||
RUN git clone --recursive https://github.com/dmlc/xgboost && \
|
||||
cd xgboost && \
|
||||
make -j4 && \
|
||||
cd python-package; python setup.py install
|
||||
|
||||
# Download code and install dependencies
|
||||
RUN cd .. && \
|
||||
git clone https://github.com/kubeflow/examples.git && \
|
||||
cd examples && \
|
||||
cd xgboost && \
|
||||
pip install -r seldon_serve/requirements.txt
|
||||
|
||||
ENTRYPOINT ["python", "examples/xgboost/housing.py"]
|
||||
|
|
@ -0,0 +1,211 @@
|
|||
# Ames housing value prediction using XGBoost on Kubeflow
|
||||
|
||||
In this example we will demonstrate how to use Kubeflow with XGBoost using the [Kaggle Ames Housing Prices prediction](https://www.kaggle.com/c/house-prices-advanced-regression-techniques/). We will do a detailed
|
||||
walk-through of how to implement, train and serve the model. You will be able to run the exact same workload on-prem and/or on any cloud provider. We will be using [Google Kubernetes Engine](https://cloud.google.com/kubernetes-engine/) to show how the end-to-end workflow runs on [Google Cloud Platform](https://cloud.google.com/).
|
||||
|
||||
# Pre-requisites
|
||||
|
||||
As a part of running this setup on Google Cloud Platform, make sure you have enabled the [Google
|
||||
Kubernetes Engine](https://cloud.google.com/kubernetes-engine/). In addition to that you will need to install
|
||||
[Docker](https://docs.docker.com/install/) and [gcloud](https://cloud.google.com/sdk/downloads). Note that this setup can run on-prem and on any cloud provider, but here we will demonstrate GCP cloud option. Finally, follow the [instructions](https://www.kubeflow.org/docs/started/getting-started-gke/) to create a GKE cluster.
|
||||
|
||||
# Steps
|
||||
* [Kubeflow Setup](#kubeflow-setup)
|
||||
* [Data Preparation](#data-preparation)
|
||||
* [Dockerfile](#dockerfile)
|
||||
* [Model Training on GKE](#model-training-on-gke)
|
||||
* [Model Export](#model-export)
|
||||
* [Model Serving Locally](#model-serving-locally)
|
||||
* [Deploying Model to Kubernetes Cluster](#model-serving-on-gke)
|
||||
|
||||
## Kubeflow Setup
|
||||
In this part you will setup Kubeflow on an existing Kubernetes cluster. Checkout the Kubeflow [getting started guide](https://www.kubeflow.org/docs/started/getting-started/).
|
||||
|
||||
## Data Preparation
|
||||
You can download the dataset from the [Kaggle competition](https://www.kaggle.com/c/house-prices-advanced-regression-techniques/data). In order to make it convenient we have uploaded the dataset on GCS
|
||||
|
||||
```
|
||||
gs://kubeflow-examples-data/ames_dataset/
|
||||
```
|
||||
|
||||
## Dockerfile
|
||||
We have attached a Dockerfile with this repo which you can use to create a
|
||||
docker image. We have also uploaded the image to gcr.io, which you can use to
|
||||
directly download the image.
|
||||
|
||||
```
|
||||
IMAGE_NAME=ames-housing
|
||||
VERSION=v1
|
||||
```
|
||||
|
||||
Use `gcloud` command to get the GCP project
|
||||
|
||||
```
|
||||
PROJECT_ID=`gcloud config get-value project`
|
||||
```
|
||||
|
||||
Let's create a docker image from our Dockerfile
|
||||
|
||||
```
|
||||
docker build -t gcr.io/$PROJECT_ID/${IMAGE_NAME}:${VERSION} .
|
||||
```
|
||||
|
||||
Once the above command is successful you should be able to see the docker
|
||||
images on your local machine by running `docker images`. Next we will upload the image to
|
||||
[Google Container Registry](https://cloud.google.com/container-registry/)
|
||||
|
||||
```
|
||||
gcloud auth configure-docker
|
||||
docker push gcr.io/${PROJECT_ID}/${IMAGE_NAME}:${VERSION}
|
||||
```
|
||||
|
||||
A public copy is available at `gcr.io/kubeflow-examples/ames-housing:v1`.
|
||||
|
||||
|
||||
## Model training on GKE
|
||||
In this section we will run the above docker container on a [Google Kubernetes Engine](https://cloud.google.com/kubernetes-engine/). There are two steps to perform the training
|
||||
|
||||
* Create a GKE cluster
|
||||
|
||||
* Create a Persistent Volume
|
||||
* Follow the instructions [here](https://kubernetes.io/docs/tasks/configure-pod-container/configure-persistent-volume-storage/). You will need to run the following `kubectl create` commands in order to get the `claim` attached to the `pod`.
|
||||
|
||||
```
|
||||
kubectl create -f py-volume.yaml
|
||||
kubectl create -f py-claim.yaml
|
||||
```
|
||||
|
||||
|
||||
* Run docker container on GKE
|
||||
* Use the `kubectl` command to run the image on GKE
|
||||
|
||||
```
|
||||
kubectl create -f py-pod.yaml
|
||||
```
|
||||
|
||||
Once the above command finishes you will have an XGBoost model available at Persistent Volume `/mnt/xgboost/housing.dat`
|
||||
|
||||
## Model Export
|
||||
The model is exported to the location `/tmp/ames/housing.dat`. We will use [Seldon Core](https://github.com/SeldonIO/seldon-core/) to serve the model asset. In order to make the model servable we have created `xgboost/seldon_serve` with the following assets
|
||||
|
||||
* `HousingServe.py`
|
||||
* `housing.dat`
|
||||
* `requirements.txt`
|
||||
|
||||
## Model Serving Locally
|
||||
We are going to use [seldon-core](https://github.com/SeldonIO/seldon-core/) to serve the model. [HousingServe.py](seldon_serve/HousingServe.py) contains the code to serve the model. Run the following command to create a microservice
|
||||
|
||||
```
|
||||
docker run -v $(pwd):/seldon_serve seldonio/core-python-wrapper:0.7 /seldon_serve HousingServe 0.1 gcr.io --base-image=python:3.6 --image-name=${PROJECT_ID}/housingserve
|
||||
```
|
||||
|
||||
Let's build the seldon-core microservice image. You can find seldon core model wrapping details [here](https://github.com/SeldonIO/seldon-core/blob/master/docs/wrappers/python.md).
|
||||
|
||||
```
|
||||
cd build
|
||||
./build_image.sh
|
||||
```
|
||||
|
||||
You should see the docker image locally `gcr.io/cloudmlplat/housingserve` which can be run locally to serve the model. Before running the image locally push it to `gcr.io`
|
||||
|
||||
```
|
||||
gcloud auth configure-docker
|
||||
docker push gcr.io/${PROJECT_ID}/housingserve:0.1
|
||||
```
|
||||
|
||||
Let's run the docker image now
|
||||
|
||||
```
|
||||
docker run -p 5000:5000 gcr.io/cloudmlplat/housingserve:0.1
|
||||
```
|
||||
|
||||
Now you are ready to send requests on `localhost:5000`
|
||||
|
||||
```
|
||||
curl -H "Content-Type: application/x-www-form-urlencoded" -d 'json={"data":{"tensor":{"shape":[1,37],"values":[1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23,24,25,26,27,28,29,30,31,32,33,34,35,36,37]}}}' http://localhost:5000/predict
|
||||
```
|
||||
|
||||
```
|
||||
{
|
||||
"data": {
|
||||
"names": [
|
||||
"t:0",
|
||||
"t:1"
|
||||
],
|
||||
"tensor": {
|
||||
"shape": [
|
||||
1,
|
||||
2
|
||||
],
|
||||
"values": [
|
||||
97522.359375,
|
||||
97522.359375
|
||||
]
|
||||
}
|
||||
}
|
||||
}
|
||||
```
|
||||
|
||||
## Model serving on GKE
|
||||
One of the amazing features of Kubernetes is that you can run it anywhere i.e., local, on-prem and cloud. We will show you how to run your code on Google Kubernetes Engine. First off, start a GKE cluster.
|
||||
|
||||
Deploy Seldon core to your GKE cluster by following the instructions in the Deploy Seldon Core section [here](https://github.com/kubeflow/examples/blob/fb2fb26f710f7c03996f08d81607f5ebf7d5af09/github_issue_summarization/serving_the_model.md#deploy-seldon-core). Once everything is successful you can verify it using `kubectl get pods -n${NAMESPACE}`.
|
||||
|
||||
```
|
||||
NAME READY STATUS RESTARTS AGE
|
||||
ambassador-849fb9c8c5-5kx6l 2/2 Running 0 16m
|
||||
ambassador-849fb9c8c5-pww4j 2/2 Running 0 16m
|
||||
ambassador-849fb9c8c5-zn6gl 2/2 Running 0 16m
|
||||
redis-75c969d887-fjqt8 1/1 Running 0 30s
|
||||
seldon-cluster-manager-6c78b7d6c9-6qhtg 1/1 Running 0 30s
|
||||
spartakus-volunteer-66cc8ccd5b-9f8tw 1/1 Running 0 16m
|
||||
tf-hub-0 1/1 Running 0 16m
|
||||
tf-job-dashboard-7b57c549c8-bfpp8 1/1 Running 0 16m
|
||||
tf-job-operator-594d8c7ddd-lqn8r 1/1 Running 0 16m
|
||||
```
|
||||
Deploy the XGBoost model
|
||||
|
||||
```
|
||||
ks generate seldon-serve-simple xgboost-ames \
|
||||
--name=xgboost-ames \
|
||||
--image=gcr.io/cloudmlplat/housingserve:0.1 \
|
||||
--namespace=${NAMESPACE} \
|
||||
--replicas=1
|
||||
|
||||
ks apply ${KF_ENV} -c xgboost-ames
|
||||
```
|
||||
|
||||
## Sample request and response
|
||||
Seldon Core uses ambassador to route its requests. To send requests to the model, you can port-forward the ambassador container locally:
|
||||
|
||||
```
|
||||
kubectl port-forward $(kubectl get pods -n ${NAMESPACE} -l service=ambassador -o jsonpath='{.items[0].metadata.name}') -n ${NAMESPACE} 8080:80
|
||||
|
||||
```
|
||||
|
||||
Now you are ready to send requests on `localhost:8080`
|
||||
|
||||
```
|
||||
curl -H "Content-Type: application/x-www-form-urlencoded" -d 'json={"data":{"tensor":{"shape":[1,37],"values":[1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23,24,25,26,27,28,29,30,31,32,33,34,35,36,37]}}}' http://localhost:8080/predict
|
||||
```
|
||||
|
||||
```
|
||||
{
|
||||
"data": {
|
||||
"names": [
|
||||
"t:0",
|
||||
"t:1"
|
||||
],
|
||||
"tensor": {
|
||||
"shape": [
|
||||
1,
|
||||
2
|
||||
],
|
||||
"values": [
|
||||
97522.359375,
|
||||
97522.359375
|
||||
]
|
||||
}
|
||||
}
|
||||
}
|
||||
```
|
||||
|
|
@ -0,0 +1,121 @@
|
|||
# Copyright 2018 Google Inc. All Rights Reserved.
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# Unless required by applicable law or agreed to in writing, software
|
||||
# distributed under the License is distributed on an "AS IS" BASIS,
|
||||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
|
||||
import argparse
|
||||
import joblib
|
||||
import numpy as np
|
||||
import pandas as pd
|
||||
from sklearn.preprocessing import Imputer
|
||||
from sklearn.model_selection import train_test_split
|
||||
from xgboost import XGBRegressor
|
||||
from sklearn.metrics import mean_absolute_error
|
||||
|
||||
def read_input(file_name, test_size=0.25):
|
||||
"""Read input data and split it into train and test."""
|
||||
data = pd.read_csv(file_name[0])
|
||||
data.dropna(axis=0, subset=['SalePrice'], inplace=True)
|
||||
|
||||
y = data.SalePrice
|
||||
X = data.drop(['SalePrice'], axis=1).select_dtypes(exclude=['object'])
|
||||
|
||||
train_X, test_X, train_y, test_y = train_test_split(X.values,
|
||||
y.values,
|
||||
test_size=test_size)
|
||||
|
||||
imputer = Imputer()
|
||||
train_X = imputer.fit_transform(train_X)
|
||||
test_X = imputer.transform(test_X)
|
||||
|
||||
return (train_X, train_y), (test_X, test_y)
|
||||
|
||||
def train_model(train_X,
|
||||
train_y,
|
||||
test_X,
|
||||
test_y,
|
||||
n_estimators,
|
||||
learning_rate):
|
||||
"""Train the model using XGBRegressor."""
|
||||
model = XGBRegressor(n_estimators=n_estimators,
|
||||
learning_rate=learning_rate)
|
||||
|
||||
model.fit(train_X,
|
||||
train_y,
|
||||
early_stopping_rounds=40,
|
||||
eval_set=[(test_X, test_y)])
|
||||
|
||||
print("Best RMSE on eval: {:.2f} with {} rounds".format(
|
||||
model.best_score,
|
||||
model.best_iteration+1))
|
||||
return model
|
||||
|
||||
def eval_model(model, test_X, test_y):
|
||||
"""Evaluate the model performance."""
|
||||
predictions = model.predict(test_X)
|
||||
print
|
||||
print("MAE on test: {:.2f}".format(mean_absolute_error(predictions, test_y)))
|
||||
|
||||
def save_model(model, model_file):
|
||||
"""Save XGBoost model for serving."""
|
||||
joblib.dump(model, model_file)
|
||||
print("Model export success {}".format(model_file))
|
||||
|
||||
def main(args):
|
||||
(train_X, train_y), (test_X, test_y) = read_input(args.train_input)
|
||||
model = train_model(train_X,
|
||||
train_y,
|
||||
test_X,
|
||||
test_y,
|
||||
args.n_estimators,
|
||||
args.learning_rate)
|
||||
|
||||
eval_model(model, test_X, test_y)
|
||||
save_model(model, args.model_file)
|
||||
|
||||
if __name__=='__main__':
|
||||
parser = argparse.ArgumentParser()
|
||||
parser.add_argument(
|
||||
'--train-input',
|
||||
help="Input training file",
|
||||
nargs='+',
|
||||
required=True
|
||||
)
|
||||
parser.add_argument(
|
||||
'--n-estimators',
|
||||
help='Number of trees in the model',
|
||||
type=int,
|
||||
default=1000
|
||||
)
|
||||
parser.add_argument(
|
||||
'--learning-rate',
|
||||
help='Learning rate for the model',
|
||||
default=0.1
|
||||
)
|
||||
parser.add_argument(
|
||||
'--model-file',
|
||||
help='Model file location for XGBoost',
|
||||
required=True
|
||||
)
|
||||
parser.add_argument(
|
||||
'--test-size',
|
||||
help='Fraction of training data to be reserved for test',
|
||||
default=0.25
|
||||
)
|
||||
parser.add_argument(
|
||||
'--early-stopping-rounds',
|
||||
help='XGBoost argument for stopping early',
|
||||
default=50
|
||||
)
|
||||
|
||||
args = parser.parse_args()
|
||||
main(args)
|
||||
|
|
@ -0,0 +1,11 @@
|
|||
kind: PersistentVolumeClaim
|
||||
apiVersion: v1
|
||||
metadata:
|
||||
name: claim
|
||||
spec:
|
||||
storageClassName: manual
|
||||
accessModes:
|
||||
- ReadWriteOnce
|
||||
resources:
|
||||
requests:
|
||||
storage: 3Gi
|
||||
|
|
@ -0,0 +1,30 @@
|
|||
apiVersion: v1
|
||||
kind: Pod
|
||||
|
||||
metadata:
|
||||
name: xgboost
|
||||
labels:
|
||||
team: platform
|
||||
|
||||
spec:
|
||||
containers:
|
||||
- name: xgboost
|
||||
image: gcr.io/cloudmlplat/ames-housing:v1
|
||||
volumeMounts:
|
||||
- mountPath: "/mnt/xgboost"
|
||||
name: datadir
|
||||
args:
|
||||
- --train-input
|
||||
- /mnt/xgboost/ames_dataset/train.csv
|
||||
- --model-file
|
||||
- /mnt/xgboost/housing.dat
|
||||
- --learning-rate
|
||||
- "0.1"
|
||||
- --n-estimators
|
||||
- "30000"
|
||||
- --early-stopping-rounds
|
||||
- "50"
|
||||
volumes:
|
||||
- name: datadir
|
||||
persistentVolumeClaim:
|
||||
claimName: claim
|
||||
|
|
@ -0,0 +1,14 @@
|
|||
kind: PersistentVolume
|
||||
apiVersion: v1
|
||||
metadata:
|
||||
name: vol
|
||||
labels:
|
||||
type: local
|
||||
spec:
|
||||
storageClassName: manual
|
||||
capacity:
|
||||
storage: 10Gi
|
||||
accessModes:
|
||||
- ReadWriteOnce
|
||||
hostPath:
|
||||
path: "/mnt/stateful_partition/data2/"
|
||||
|
|
@ -0,0 +1,37 @@
|
|||
# Copyright 2018 Google Inc. All Rights Reserved.
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# Unless required by applicable law or agreed to in writing, software
|
||||
# distributed under the License is distributed on an "AS IS" BASIS,
|
||||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
|
||||
import joblib
|
||||
import numpy as np
|
||||
|
||||
class HousingServe(object):
|
||||
def __init__(self, model_file='housing.dat'):
|
||||
"""Load the housing model using joblib."""
|
||||
self.model = joblib.load(model_file)
|
||||
|
||||
def predict(self, X, feature_names):
|
||||
"""Predict using the model for given ndarray."""
|
||||
prediction = self.model.predict(data=X)
|
||||
|
||||
# added this temporarily to keep seldon happy
|
||||
# TODO: Fix https://github.com/SeldonIO/seldon-core/blob/master/wrappers/python/model_microservice.py#L55
|
||||
return [[prediction.item(0), prediction.item(0)]]
|
||||
|
||||
def sample_test(self):
|
||||
"""Generate a random sample feature."""
|
||||
return np.ndarray([1, 37])
|
||||
|
||||
if __name__=='__main__':
|
||||
serve = HousingServe()
|
||||
print(serve.predict(serve.sample_test(), None))
|
||||
Binary file not shown.
|
|
@ -0,0 +1,7 @@
|
|||
joblib==0.12.0
|
||||
numpy==1.14.4
|
||||
pandas==0.23.0
|
||||
scikit-learn==0.19.1
|
||||
scipy==1.1.0
|
||||
sklearn==0.0
|
||||
xgboost==0.72.1
|
||||
Loading…
Reference in New Issue